Oops, we made a net

September 12, 2018

The social network. A model we take for granted… It began as a simple set of links, human to human, ‘x knows y’. At best somewhat directed. In that form it’s not quite innocent—you can tell a lot about people from who they know. But it’s relatively powerless.

Add communication. X follows Y, Y posts and X sees it. Now we have a simple forward-propagation network. It may have loops, but content moves slowly and mutates heavily. Everyone has a nice day, aside from cussing at the bad MySpace page formats. Except for a few refresh obsessives, most of us go on our day unobstructed.

Add comment threads / replies. This goes two ways. Either they’re second class objects, leading to forum or livejournal style posts-with-discussion, or they’re first class objects leading to the twitter or mastodon situation. But the maliability of the generated comments gives these a flexible effect.

Add likes / favorites. Now the trouble begins. They’re fast and cheap and they do one thing the simple forward-propagation network didn’t. They quickly and succinctly back-propagate. You know whether you got likes. You soon realize why.

Add boosts / shares. And the nodes of your back-propagating perceptron now have a convenient means of both acting as filters and getting around the single greatest weakness of the typical model, namely exponential loss of back-propagation effectiveness with increase in node depth, since likes / favorites hit straight to the source and boosts / shares interact with the connection formation mechanisms.

This is a fully functional neural network. With entire human minds for nodes. Operating at the speed of your keyboard.

Add hashtags. Now the nodes connect by a layer orthogonal to their pre-existing mechanisms, allowing the rapid re-formation of connective clusters. As news events ebb and flow, the remains of these constructs will come to dominate the connective pattern. This selects for collective excitation patterns eminating from a few key players. You now have a sensory apparatus. Users sort themselves by chosen distance from various sensory clusters, as outward propagators act as filter nodes.

Add federation. By means of local and fedi timelines, a concept of locality is reclaimed which may not correspond to the sensory cluster patterns. However, over time it increasingly seems to. These provide broad arrays of lower activation input (one doesn’t always look at fedi unless bored) in addition to the high activation personal timeline. Something akin to a regional mood is achieved.

Add celebrities. Which here I define as anyone posessing a hefty skew of followers:follows, or who is unduly prone to being boosted / shared. With functioning sensory clusters, a single node is no longer recognized as a good source of information to have such a broad effect. Allergic responses follow, attempting to dampen that node, until rejection becomes inevitable. No aspect of this system beyond the nodal scale retains awareness that its effects are upon humans.

Ebbs and flows give way to flash mobs and consternation as we wonder what went wrong. Nothing went wrong. Something went inhuman / ahuman. We did. The neurons don’t know what the brain is pondering, and we’re the neurons. Shit posting and flash mobs, meme wars and refresh addiction, are all just part of being part of something. A pity we don’t quite know what, because we’re going to have to square with it.

Maybe this is something humans have always had the drive to do, if not the ability at this scale. But it’s more than simply communication. It’s a computationally complete AI built on top of a network of natural intelligences. And we have no idea what it’s up to.

Oops, we made a net - September 12, 2018 - Kim Reece